Criteo Attribution Modeling for Bidding Dataset

By: Criteo Research / 10 Oct 2017

We recently published a paper on Attribution Modelling for Bidding at TargetAd and AdKDD 2017.  As usual when we publish research work on Criteo related data, we publicly release the dataset used for the experiments in the paper.

Compared to other online advertising datasets such as CTR prediction dataset or Conversion Modelling datasets, this one contains attribution data, i.e if the conversions are attributed to Criteo or not by the advertisers. This important information opens a new set of applications in the area of real-time bidding and conversion modeling in online advertising.

In our paper we show that using a simple attribution model in the bidder can significantly improve bidding performance, providing better ROI for the advertiser and reduced ad exposure for the user,  compared to baseline last-click bidder.

Content of this dataset

This dataset includes following files:

  • criteo_attribution_dataset.tsv.gz: the dataset itself (623M compressed)
  • Experiments.ipynb: ipython notebook with code and utilities to reproduce the results in the paper. Can also be used as a starting point for further research on this data. It requires python 3.* and standard scientific libraries such as pandas, numpy and sklearn.

Data description

This dataset represents a sample of 30 days of Criteo live traffic data. Each line corresponds to one impression (a banner) that was displayed to a user. For each banner we have detailed information about the context, if it was clicked, if it led to a conversion and if it led to a conversion that was attributed to Criteo or not. Data has been sub-sampled and anonymized so as not to disclose proprietary elements.

Here is a detailed description of the fields (they are tab-separated in the file):

  • timestamp: timestamp of the impression (starting from 0 for the first impression). The dataset is sorted according to timestamp.
  • uid a unique user identifier
  • campaign a unique identifier for the campaign
  • conversion 1 if there was a conversion in the 30 days after the impression (independently of whether this impression was last click or not)
  • conversion_timestamp the timestamp of the conversion or -1 if no conversion was observed
  • conversion_id a unique identifier for each conversion (so that timelines can be reconstructed if needed). -1 if there was no conversion
  • attribution 1 if the conversion was attributed to Criteo, 0 otherwise
  • click 1 if the impression was clicked, 0 otherwise
  • click_pos the position of the click before a conversion (0 for first-click)
  • click_nb number of clicks. More than 1 if there was several clicks before a conversion
  • cost the price paid by Criteo for this display (disclaimer: not the real price, only a transformed version of it)
  • cpo the cost-per-order in case of attributed conversion (disclaimer: not the real price, only a transformed version of it)
  • time_since_last_click the time since the last click (in s) for the given impression
  • cat[1-9] contextual features associated to the display. Can be used to learn the click/conversion models. We do not disclose the meaning of these features but it is not relevant for this study. Each column is a categorical variable. In the experiments, they are mapped to a fixed dimensionality space using the Hashing Trick (see paper for reference).

Key figures

  • 2,4Gb uncompressed
  • 16.5M impressions
  • 45K conversions
  • 700 campaigns


This dataset can be used in a large scope of applications related to Real-Time-Bidding, including but not limited to:

  • Attribution modeling: rule based, model based, etc…
  • Conversion modeling in display advertising: the data includes cost and value used for computing Utility metrics.
  • Offline metrics for real-time bidding


This dataset is released along with following paper:

“Attribution Modeling Increases Efficiency of Bidding in Display Advertising”
Eustache Diemert*, Julien Meynet* (Criteo Research), Damien Lefortier (Facebook), Pierre Galland (Criteo) *authors contributed equally

published in “2017 AdKDD & TargetAd Workshop, in conjunction with The 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2017) ”

When using this dataset, please cite the paper with following bibtex (final ACM bibtex coming soon):

author = {{Diemert Eustache, Meynet Julien} and Galland, Pierre and Lefortier, Damien},
title={Attribution Modeling Increases Efficiency of Bidding in Display Advertising},
publisher = {ACM},
pages={To appear},
booktitle = {Proceedings of the AdKDD and TargetAd Workshop, KDD, Halifax, NS, Canada, August, 14, 2017},
year = {2017}

We would love to hear from you if use this data or plan to use it. Refer to the Contact section below.


For any question, feel free to contact:

Download instructions

Post written by: 

Julien Meynet

Senior Research Scientist at Criteo.